AI & Work

Decide what finished looks like before you ask AI for anything

Say a colleague asks an AI tool for "a summary of the client call". The tool produces a page. It is fluent and organised, and nobody can use it, because the colleague wanted each decision, who owns it and by when, and the tool was never told that.

Start a step before the prompt: decide what the finished output looks like. In my AI enablement work I call this output definition. It is easy to skip, because the picture of the result feels obvious to the person who already has it in mind.

Five lines before the prompt

Before you ask for anything that matters, write these five lines.

1. Who reads it. A named role, with what they already know. "The client's finance head, who was not on the call" calls for a different output from "my team, who were".

2. What they will do with it. Approve something, decide something, forward it, file it. If you cannot name the action, ask whether the output is needed at all.

3. Shape and length. Three bullets, one paragraph, a table with these columns, a word limit. Be specific enough that two people would produce the same shape.

4. What must be in it. The two or three items it fails without. For the call summary: each decision, its owner, the date.

5. How you will check it. One question you will ask of the output before you use it. Is every number traceable to the source material? Would the reader know what to do next? Did it add anything the material did not contain?

Lines one to four go into the prompt almost as written. Line five stays with you as your test, and it makes the checking quick.

What it does for a team

Ten people using AI for client summaries carry ten private ideas of a good one, and the reviewer has to reconcile them. Agree the five lines once for the outputs your team produces most, keep them where everyone can find them, and review against them. Reviews then focus on the content, and new joiners can read the standard on day one.

The five lines also give a team a way to evaluate its use of AI. Once a team has written down what good output is for a given task, it can check how often the tool gets there and where people still step in.

The other half of the job is the context you give the tool. I wrote about that in How AI Can Help Your Operations, If You Stop Treating It Like a Mind Reader.

Teams that want to build this habit together, with their own tools and content, can do it in an AI Content Ops programme, which I co-founded.

Take the AI output your team relies on every week. Who would write its five lines, and would everyone else agree with them?

Send feedback

Goes to my inbox. Nothing is published.

All posts